[
https://issues.apache.org/jira/browse/FLINK-5588?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15866685#comment-15866685
]
Stavros Kontopoulos edited comment on FLINK-5588 at 2/15/17 2:09 AM:
---------------------------------------------------------------------
[~till.rohrmann] Have already implemented the Normalizer... need to check
floating arithmetic for the UnitScaler because the sum might lead to overflow,
so have to find a proper algo.
Thinking of dividing with Xmax to avoid overflow and use
https://en.wikipedia.org/wiki/Kahan_summation_algorithm for the sum of many
small numbers.
Standard scaler uses this algo:
http://www.cs.yale.edu/publications/techreports/tr222.pdf for variance on big
vectors, need something similar.
was (Author: skonto):
[~till.rohrmann] Have already implemented the Normalizer... need to check
floating arithmetic for the UnitScaler because the sum might lead to overflow,
so have to find a proper algo.
Thinking of dividing with Xmax to avoid overflow and use
https://en.wikipedia.org/wiki/Kahan_summation_algorithm for the sum of many
small numbers.
Standard scaler uses this algo:
http://www.cs.yale.edu/publications/techreports/tr222.pdf for variance on big
vectors.
> Add a unit scaler based on different norms
> ------------------------------------------
>
> Key: FLINK-5588
> URL: https://issues.apache.org/jira/browse/FLINK-5588
> Project: Flink
> Issue Type: New Feature
> Components: Machine Learning Library
> Reporter: Stavros Kontopoulos
> Assignee: Stavros Kontopoulos
> Priority: Minor
>
> So far ML has two scalers: min-max and the standard scaler.
> A third one frequently used, is the scaler to unit.
> We could implement a transformer for this type of scaling for different norms
> available to the user.
> I will make a separate class for the Normalization per sample procedure by
> using the Transformer API because it is easy to add
> it, fit method does nothing in this case.
> Scikit-learn has also some calls available outside the Transform API, we
> might want add that in the future.
> These calls work on any axis but they are not re-usable in a pipeline [4]
> Right now the existing scalers in Flink ML support per feature normalization
> by using the Transformer API.
> Resources
> [1] https://en.wikipedia.org/wiki/Feature_scaling
> [2]
> http://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.scale.html
> [3] https://spark.apache.org/docs/2.1.0/mllib-feature-extraction.html
> [4] http://scikit-learn.org/stable/modules/preprocessing.html
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
